Using machine learning to aid treatment decision and risk assessment for severe three-vessel coronary artery disease.

BACKGROUND: Three-vessel disease (TVD) with a SYNergy between PCI with TAXus and cardiac surgery (SYNTAX) score of ≥ 23 is one of the most severe types of coronary artery disease. We aimed to take advantage of machine learning to help in decision-making and prognostic evaluation in such patients.
METHODS: We analyzed 3786 patients who had TVD with a SYNTAX score of ≥ 23, had no history of previous revascularization, and underwent either coronary artery bypass grafting (CABG) or percutaneous coronary intervention (PCI) after enrollment. The patients were randomly assigned to a training group and testing group. The C4.5 decision tree algorithm was applied in the training group, and all-cause death after a median follow-up of 6.6 years was regarded as the class label.
RESULTS: The decision tree algorithm selected age and left ventricular end-diastolic diameter (LVEDD) as splitting features and divided the patients into three subgroups: subgroup 1 (age of ≤ 67 years and LVEDD of ≤ 53 mm), subgroup 2 (age of ≤ 67 years and LVEDD of > 53 mm), and subgroup 3 (age of > 67 years). PCI conferred a patient survival benefit over CABG in subgroup 2. There was no significant difference in the risk of all-cause death between PCI and CABG in subgroup 1 and subgroup 3 in both the training data and testing data. Among the total study population, the multivariable analysis revealed significant differences in the risk of all-cause death among patients in three subgroups.
CONCLUSIONS: The combination of age and LVEDD identified by machine learning can contribute to decision-making and risk assessment of death in patients with severe TVD. The present results suggest that PCI is a better choice for young patients with severe TVD characterized by left ventricular dilation. Copyright and License information: Journal of Geriatric Cardiology 2022.

Coronary artery disease (CAD) is the leading cause of death and disability worldwide.[ Three-vessel disease (TVD) is the most severe form of CAD and is characterized by significant stenosis in all three major coronary arteries. The application of myocardial revascularization techniques, including coronary artery bypass grafting (CABG) and percutaneous coronary intervention (PCI), has significantly improved the clinical outcomes of patients with severe CAD. CABG has traditionally been the standard therapy for complex coronary lesions, including TVD.[ In recent years, with the advancements in PCI technology and the accumulation of operators’ experience, the incidence of periprocedural and long-term adverse events of PCI has substantially decreased, and PCI has been gradually applied in the treatment of TVD.[ Current guidelines recommend use of the SYNergy between PCI with TAXus and cardiac surgery (SYNTAX) score and diabetes status to guide the revascularization strategy for patients with TVD.[ Current practice guidelines do not recommend PCI for patients with TVD with a SYNTAX score of ≥ 23. However, using the SYNTAX score to guide clinical decision-making between PCI and CABG remains controversial, and its reasonability has been questioned since a newly published meta-analysis showed no significant association between the SYNTAX score and the comparative effectiveness of PCI and CABG.[ Moreover, the SYNTAX score is a quantitative indicator of the anatomical complexity of TVD and does not include clinical variables that may have significant effects on the patient’s prognosis. Whether some patients with specific clinical characteristics can obtain a comparable or even greater survival benefit from PCI than from CABG is unclear. Moreover, risk assessment for patients with TVD after revascularization therapy remain challenging.[

Machine learning has recently emerged as an important research method and has been successfully applied in many scientific fields, including clinical medicine.[ The decision tree algorithm, a common approach in machine learning, can handle non-linearity, heterogeneous effects, and high-dimensional features and partition a trial population into subgroups characterized by multiple simultaneous characteristics.[ In the present study involving a large cohort of patients with TVD with a SYNTAX score of ≥ 23, we employed a decision tree algorithm to generate specific subgroups, compared the long-term prognosis between patients who underwent PCI or CABG in each subgroup, and conducted a comparative analysis of the long-term prognosis between subgroups generated by machine learning. We evaluated whether machine learning can help in selecting the optimal revascularization method and assessing risk in patients with severe TVD.

PATIENTS AND METHODS

Patient Selection

This was a prospective, single-center study. All patients were from a large cohort of 8943 patients with TVD.[ The patients were consecutively enrolled in Fuwai Hospital (Beijing, China) from April 2004 to February 2011. TVD was defined as angiographic stenosis of ≥ 50% in all three main coronary arteries, including the left anterior descending, left circumflex, and right coronary arteries. The SYNTAX score was calculated using an online calculator ( http://www.syntaxscore.com) by an experienced interventional cardiologist who was blinded to the clinical data. The study population comprised 3786 patients with TVD and a SYNTAX score of ≥ 23, with no history of previous revascularization therapy, and who underwent either CABG or PCI after enrollment (Supplemental Figure 1). There were no predetermined exclusion criteria. The choice of PCI or CABG mainly followed the contemporary practice guidelines, judgment by the heart team, and patients’ preference. All patients’ clinical data were collected at baseline and entered into a database by independent clinical research coordinators.

This study complied with the principles of the Declaration of Helsinki and was approved by the Ethics Committee of Fuwai Hospital. All patients provided written informed consent.

Outcomes

After enrollment, the patients were followed up according to the study protocol.[ Outcome data were obtained by telephone interview, follow-up letter, or clinic visit. The last follow-up was completed in 2016. All events were carefully checked and verified by an independent group of clinical physicians. Investigator training, blinded questionnaire filling, and telephone recording were performed to obtain high-quality data. The primary endpoint was all-cause death. The secondary endpoints were major adverse cardiac and cerebrovascular events (MACCE), which were a composite of all-cause death, myocardial infarction and stroke. All patients who were enrolled in the analysis completed at least one follow-up, and patients who were lost to follow-up were calculated according to censored data.

Machine Learning

We used the C4.5 decision tree algorithm, a supervised machine learning method, to build splitting rules from patients with a list of features and a class label. The analysis process is shown in a flow diagram in the Graphical Abstract and Supplemental Figure 1. Specifically, about 60% of the total participants were randomly assigned to the training group, and the remaining participants were assigned to the testing group. In the training group, patients undergoing CABG were used to generate splitting rules because CABG was the standard treatment for patients with TVD with a SYNTAX score of ≥ 23 according to the current practice guidelines at the time of study. Eighty clinical parameters were included in the data set as features and the parameters with > 10% missing values were excluded from analysis (Supplemental Table 1 and Supplemental Table 2). The presence or absence of all-cause death was regarded as a class label for each patient. The C4.5 algorithm uses the information gain ratio, which is an information theory metric, to select the informative feature to divide the dataset into several subsets with different mortality rates. By calculating the information gain ratio of each feature in the data set, the algorithm selects the feature and corresponding cut-point with the maximum information gain ratio as a splitting feature and recursively constructs a tree-like model. The stopping criterion was ≥ 300 patients in each subset. Finally, the patients were split into subsets such that the data in each subset had a higher degree of aggregation. The same splitting rules were then applied to patients undergoing PCI to generate concordant subsets. All algorithms were implemented on the Waikato Environment for Knowledge Analysis (WEKA) platform.

Statistical Analysis

Continuous variables are presented as mean ± SD or median (interquartile range), and categorical variables are presented as number with frequency. Differences in participant characteristics were compared by the chi-square test for categorical variables and by Student’s t test or the Mann–Whitney U test for continuous variables. The cumulative survival rate was calculated using the Kaplan–Meier method and difference among treatment groups were assessed by the log-rank test. Univariable and multivariable Cox proportional hazard regressions were performed to calculate the hazard ratio (HR) and 95% confidence interval (CI) and evaluate the associations between the treatment method and clinical outcomes. The multivariable model was adjusted for the following potential confounders using an all-enter way: age, sex, body mass index, peripheral artery disease, chronic kidney disease, left main involvement, left ventricular ejection fraction, creatine clearance, and SYNTAX score. All statistical analyses were two-sided, and P< 0.05 was considered statistically significant. Statistical analyses were performed using SPSS version 24.0 (IBM Corporation, Armonk, NY, USA).

RESULTS

Patients’ Clinical Characteristics

In total, 3786 patients with TVD with a SYNTAX score of ≥ 23 were included in the study. Their baseline characteristics are shown in Table 1. The patients’ mean age was 60.9 years, and 81.0% were male. PCI was performed in 1660 (43.8%) patients and CABG in 2126 (56.2%). Patients undergoing CABG tended to be slightly older and had more complex lesions as reflected by a significantly higher SYNTAX score (P < 0.001).

Table 1

Clinical characteristics of the study population.

	All (n = 3786)	Patients with PCI (n = 1660)	Patients with CABG (n = 2126)	P-value
Values are presented as mean ± SD or n (%). BMI: body mass index; CABG: coronary artery bypass grafting; CrCl: creatinine clearance; CKD: chronic kidney disease; COPD: chronic obstructive pulmonary disease; LVEDD: left ventricular end-diastolic diameter; LVEF: left ventricular ejection fraction; NT-proBNP: N-terminal pro-B-type natriuretic peptide; PAD: peripheral artery disease; PCI: percutaneous coronary intervention. a Calculated using the Cockcroft and Gault formula.
Age, yrs	60.9 ± 9.6	60.5 ± 10.6	61.2 ± 8.8	0.032
Male	3067 (81.0%)	1314 (79.2%)	1753 (82.5%)	0.010
BMI, kg/m2	25. 8 ± 3.0	25.9 ± 3.0	25.7 ± 3.0	0.071
Diabetes	1330 (35.1%)	603 (36.3%)	727 (34.2%)	0.173
COPD	38 (1.0%)	21 (1.3%)	17 (0.8%)	0.154
PAD	328 (8.7%)	73 (4.4%)	255 (12.0%)	< 0.001
CKD	25 (0.7%)	6 (0.4%)	19 (0.9%)	0.045
Smoking history	2093 (55.3%)	926 (55.8%)	1167 (54.9%)	0.584
Syntax score	31.8 ± 8.6	29.3 ± 5.5	33.7 ± 10.0	< 0.001
Left main involvement	2412 (63.7%)	1281 (77.2%)	1131 (53.2%)	< 0.001
LVEF	58.9% ± 8.9%	59.4% ± 8.6%	58.5% ± 9.1%	0.009
LVEDD, mm	50.1 ± 5.8	50.0 ± 5.4	50.2 ± 6.1	0.759
NT-proBNP, pmol/L	815.91 ± 677.77	813.61 ± 635.81	817.73 ± 709.35	0.851
aCrCl, mL/min	85.73 ± 25.93	87.33 ± 27.56	94.47 ± 24.51	0.003

Long-term Outcomes of PCI and CABG in Total Population

During the median follow-up of 6.6 years, 442 (11.7%) patients experienced all-cause death and 856 (22.6%) patients developed MACCE. Kaplan-Meier analysis showed that the cumulative mortality rates in the PCI and CABG groups were 12.6% and 11.0%, respectively (log-rank test, P = 0.360) and that the cumulative MACCE rates in PCI and CABG groups were 24.0% and 21.5%, respectively (log-rank test, P = 0.404), indicating no significant difference (Figure 1).

Figure 1

Comparison of long-term prognosis of PCI and CABG groups in total study population

Decision Tree Model

In total, 2162 (57.1%) patients were randomly assigned to the training group. The decision tree algorithm split patients undergoing CABG into three subgroups by age and left ventricular end-diastolic diameter (LVEDD) (see Graphical Abstract). Subgroup 1 was defined by an age of ≤ 67 years and LVEDD of ≤ 53 mm; subgroup 2, age of ≤ 67 years and LVEDD of > 53 mm; and subgroup 3, age of > 67 years. Patients undergoing PCI were also divided into three subgroups based on the same splitting rules, and patients in the corresponding subgroups of the PCI group and CABG group came together to form three subgroups.

Comparison of Long-term Prognosis of PCI and CABG in Each Subgroup of Training Data and Testing Data

In the training data, 1130 (52.3%) patients were classified into subgroup 1 (PCI, 44.4%; CABG, 55.6%), and 603 (27.9%) patients were classified into subgroup 3 (PCI, 46.3%; CABG, 53.7%) (Table 2). There was no significant difference in the incidence of all-cause death or MACCE between these two subgroups (Figure 2 and Supplemental Figure 2). In contrast, 429 (19.8%) patients were classified into subgroup 2 (PCI, 38.9%; CABG, 61.1%). PCI resulted in a significantly lower all-cause mortality rate than CABG (8.4% vs. 15.3%, P = 0.049), but there was no significant difference in the incidence of MACCE (25.1% vs. 24.8%, P = 0.799). Moreover, the mortality rate after CABG was only 4.8% in young patients without left ventricular dilatation (subgroup 1), but was substantially higher at 15.3% in young patients with left ventricular dilatation (subgroup 2) (Table 2). In the testing data, patients were divided into three subgroups based on same splitting rules as those used in the training data. The incidence of all-cause death was not significantly different between PCI and CABG in subgroups 1 and 3, but showed a significant difference in subgroup 2 (4.5% vs. 13.1%, respectively; P = 0.037) (Table 2 and Figure 2). Moreover, there was no significant difference in the incidence of MACCE between PCI and CABG among the three subgroups in the testing data (Supplemental Figure 2). The long-term outcomes of PCI and CABG in the three subgroups showed a similar tendency as those in the training data (Table 2 and Figure 2).

Table 2

Summarize of the incidence of adverse event in different subgroups identified by decision tree analysis.

Group	Patients with PCI	Patients with CABG	Death in PCI	Death in CABG	MACCE in PCI	MACE in CABG	MI in PCI	MI in CABG	Stroke in PCI	Stroke in CABG
Values are presented as n (%). CABG: coronary artery bypass grafting; MACCE: major adverse cardiac and cerebrovascular events; MI: myocardial infarction; PCI: percutaneous coronary intervention.
Training data
Subgroup 1	502 (44.4%)	628 (55.6%)	36 (7.2%)	30 (4.8%)	90 (17.9%)	105 (16.7%)	40 (8.0%)	14 (2.2%)	17 (3.4%)	67 (10.7%)
Subgroup 2	167 (38.9%)	262 (61.1%)	14 (8.4%)	40 (15.3%)	42 (25.1%)	65 (24.8%)	21 (12.6%)	5 (1.9%)	12 (7.2%)	22 (8.4%)
Subgroup 3	279 (46.3%)	324 (53.7%)	70 (25.1%)	62 (19.1%)	102 (36.6%)	91 (28.1%)	20 (7.2%)	6 (1.9%)	20 (7.2%)	32 (9.9%)
Testing data
Subgroup 1	393 (44.4%)	492 (55.6%)	30 (7.6%)	24 (4.9%)	72 (18.3%)	73 (14.8%)	34 (8.7%)	15 (3.0%)	14 (3.6%)	37 (7.5%)
Subgroup 2	110 (38.5%)	176 (61.5%)	5 (4.5%)	23 (13.1%)	18 (16.4%)	41 (23.3%)	6 (5.5%)	6 (3.4%)	10 (9.1%)	17 (9.7%)
Subgroup 3	209 (46.1%)	244 (53.9%)	54 (25.8%)	54 (22.1%)	74 (35.4%)	83 (34.0%)	10 (4.8%)	7 (2.9%)	15 (7.2%)	30 (12.3%)

Figure 2

Comparison of long-term prognosis of PCI and CABG in each subgroup.

Comparison of Long-term Prognosis of Three Subgroups in Total Study Population

In the training data, 1130 (52.3%) patients were classified into subgroup 1 (PCI, 44.4%; CABG, 55.6%), and 603 (27.9%) patients were classified into subgroup 3 (PCI, 46.3%; CABG, 53.7%) (Table 2). There was no significant difference in the incidence of all-cause death or MACCE between these two subgroups (Figure 2 and Supplemental Figure 2). In contrast, 429 (19.8%) patients were classified into subgroup 2 (PCI, 38.9%; CABG, 61.1%). PCI resulted in a significantly lower all-cause mortality rate than CABG (8.4% vs. 15.3%, P = 0.049), but there was no significant difference in the incidence of MACCE (25.1% vs. 24.8%, P = 0.799). Moreover, the mortality rate after CABG was only 4.8% in young patients without left ventricular dilatation (subgroup 1), but was substantially higher at 15.3% in young patients with left ventricular dilatation (subgroup 2) (Table 2). In the testing data, patients were divided into three subgroups based on same splitting rules as those used in the training data. The incidence of all-cause death was not significantly different between PCI and CABG in subgroups 1 and 3, but showed a significant difference in subgroup 2 (4.5% vs. 13.1%, respectively; P = 0.037) (Table 2 and Figure 2). Moreover, there was no significant difference in the incidence of MACCE between PCI and CABG among the three subgroups in the testing data (Supplemental Figure 2). The long-term outcomes of PCI and CABG in the three subgroups showed a similar tendency as those in the training data (Table 2 and Figure 2).

DISCUSSION

Using a machine learning approach, we found that age and LVEDD can help in selecting the revascularization method and assessing the risk of death for patients with TVD with a SYNTAX score of ≥ 23.

TVD is one of the most severe forms of CAD, and the current clinical practice guidelines recommend CABG as the only effective revascularization method for patients with TVD with a SYNTAX score of ≥ 23.[ However, the most recent 10-year follow-up results of the SYNTAX trial revealed no significant difference in all-cause mortality between PCI and CABG among patients with TVD and a SYNTAX score of 23 to 32.[ Recently published studies also questioned the reasonability of using SYNTAX score to guide decision-making between PCI and CABG.[ Moreover, PCI has the advantages of being minimally invasive, having a lower risk of iatrogenic injuries, and having a shorter hospitalization time. Identifying specific patients for whom PCI can act as an alternative or better therapy than CABG is of great clinical importance and requires further investigation.

We used the advantages of machine learning to aid the present analysis and employed the C4.5 decision tree algorithm. All available clinical parameters, including the medical history, clinical presentation, and results of auxiliary examinations, were incorporated as candidate features into the generation of subgroups to obtain a comprehensive result. Different combinations of candidate features and specific cut-points can generate different subgroups, and tens of thousands of potential combinations exist. The C4.5 decision tree algorithm can identify the optimal combination based on information entropy theory and recursive partitioning of patients into subsets with a high degree of class aggregation.[ Unlike black-box machine learning methods (e.g., neural networks and support vector machines), in which the original classification rules are unknowable, the C4.5 decision tree algorithm is a white-box machine learning method that can output the original results of the features and the selected cut-point to generate subgroups.[ Moreover, the minimum number of patients in the subgroups generated by the C4.5 decision tree algorithm can be user-controlled by setting up a stopping criterion. In the present study, the stopping criterion was ≥ 300 patients in each subset to avoid the inability to perform the subsequent statistical analysis because of a too-small patient size in each subgroup.

We found that patients with TVD with a SYNTAX score of ≥ 23 could be divided into three subgroups based on a combination of age and LVEDD. For patients aged ≤ 67 years and with significant left ventricular dilatation (LVEDD of > 53 mm), PCI conferred a long-term event-free survival benefit of all-cause death over CABG. For older patients (age of > 67 years) and young patients without significant left ventricular dilatation (age of ≤ 67 years and LVEDD of ≤ 53 mm), the mortality rate was lower in the CABG group than in the PCI group, but the difference was not significant. Previous meta-analyses also showed that patient age significantly modified the treatment effect of CABG and PCI on mortality and that CABG is a better choice for patients of older ages whereas PCI is better for patients of younger ages. [ The meta-analyses did not show a difference in survival with PCI and CABG between patients with normal and abnormal left ventricular function. Several parameters are indicators of cardiac function, such as the left ventricular ejection fraction, LVEDD, and N-terminal pro-B-type natriuretic peptide concentration. In the present study, without predefined postulates, the LVEDD was selected by machine learning to guide the decision-making process. We found that among young patients (age ≤ 67 years), those with significant left ventricular dilatation (LVEDD > 53 mm) had a worse prognosis following CABG than patients without significant left ventricular dilatation. A previous study also demonstrated that a larger heart size in both its diastolic and systolic dimensions is associated with a poorer predicted response to CABG. [ These results imply that left ventricular dilatation may impair the treatment effect of CABG. Our results will improve the decision-making process in selection of a more individualized revascularization method for patients who have TVD with a SYNTAX score of ≥ 23.

Risk assessment for patients with complex CAD after revascularization treatment remains challenging. Researchers have attempted to build a model to predict mortality risk after CABG and PCI for individual patients. One such example is the SYNTAX score II, which was developed to predict 4-year mortality after CABG or PCI by combining the original SYNTAX score and seven clinical variables for patients with complex CAD, including TVD. However, this score showed only moderate discrimination ability.[ Researchers have found that combining the SYNTAX score II with other clinical variables such as the N-terminal pro-B-type natriuretic peptide concentration, hyperuricemia, the big endothelin-1 concentration, and the white blood cell count can improve the predictive performance in patients with TVD, implying that further construction of a comprehensive predictive model that incorporates more clinical parameters than the SYNTAX score II might be of value.[ However, such a model may be too complex for clinical utility. Substantial time, resources, and effort may be further required to develop and maintain a Web-based calculator or mobile app to facilitate use of the model by clinicians. Clinicians may also need to spend a considerable amount of time using the calculator for decision making. In this study, we proposed a novel and totally different idea: dividing patients into different subgroups based on a combination of variables identified by machine learning. Our results indicated that patients in the three subgroups showed significant differences in long-term mortality. Young patients without left ventricular dilation (age of ≤ 67 years and LVEDD of ≤ 53 mm) had the best prognosis, whereas young patients with left dilation (age of ≤ 67 years and LVEDD of > 53 mm) showed a worse prognosis and older patients (age of > 67 years) had the worst prognosis.

Different from conventional statistical methods, machine learning uses automatic selection and combinations of variables and provides a definite cut-point value. For instance, age is a conventional risk marker for a poor prognosis in patients with CVD and is also one of the parameters incorporated in the SYNTAX score II.[ Patients of advanced age are thought to have a poor prognosis, but older age is a vague cut-off.[ Conventional studies only indicate by how many times the estimated death risk may increase with every 1-, 5-, or 10-year increase in age, for example, by conducting Cox regression analysis. Our results provided a specific age of 67 years as the cut-point to guide selection of the revascularization method and risk assessment in combination with other clinical variables. The same was the case for LVEDD. Although the optimal cut-point value needs to be further investigated and validated in larger-scale, multicenter studies and may vary in different populations, we have provided a possible solution to compensate for the limitations of conventional statistical methods.

LIMITATIONS

All participants were from a single tertiary center, which may have resulted in selection bias. Additionally, two variables were selected to divide the patients into three subgroups by a decision tree algorithm. However, the number of subgroups using this method is related to the number of patients. A larger patient population allows more subgroups to be generated and more refined treatment strategies to be obtained. Although we conducted internal validations by using a set of data as testing data in our cohort, we acknowledge that external validation was not yet available for the present study, and this may have affected the robustness of our results. Our findings will require validation in an independent external cohort.

CONCLUSION

Machine learning analysis classified patients who had TVD with a SYNTAX score of ≥ 23 into three subgroups by age and LVEDD, guiding the establishment of a more individualized strategy of selecting the revascularization method and contributing to risk assessment of death for patients with severe TVD. PCI may be a better option for relatively young patients with left ventricular dilation. For older patients or young patients without left ventricular dilation, PCI might be an acceptable option when the patients have surgical contraindications or cannot tolerate CABG.

FUNDINGS

This work was supported by the CAMS Innovation Fund for Medical Sciences (grant number 2016-I2M-1-002); the Beijing Municipal Natural Science Foundation (grant number 7181008), and Capital’s Funds for Health Improvement and Research (grant number 2018-2-4033).

CONFLICT OF INTERESTS

None.

(A): all-cause death; (B): MACCE. CABG: coronary artery bypass grafting; MACCE: major adverse cardiac and cerebrovascular events; PCI: percutaneous coronary intervention.

(A): Training data in subgroup 1; (B): training data in subgroup 2; (C): training data in subgroup 3; (D): testing data in subgroup 1; (E): testing data in subgroup 2; and (F) testing data in subgroup 3. CABG: coronary artery bypass grafting; PCI: percutaneous coronary intervention.

Comparison of long-term prognosis of three subgroups.

(A): Training data of all-cause death; (B): testing data of all-cause death; (C): total population of all-cause death; (D): training data of cardiovascular death; (E): testing data of cardiovascular death; and (F): total population of cardiovascular death. CABG: coronary artery bypass grafting; PCI: percutaneous coronary intervention.

Supplementary data to this article can be found online.

Click here for additional data file.